Tom Stellard <thomas.stell...@amd.com> writes: > Hi, > > I've been testing these updated compute patches all week and they look > good to me. I don't think there are any outstanding complaints, so I'll > give my ACK for merging these into master. > > Very nice work! > > -Tom Stellard > > Thanks for your comments. Does anyone else have anything to add before I merge the gallium-compute branch to master?
>
> On Tue, May 01, 2012 at 05:27:39PM +0200, Francisco Jerez wrote:
>> Define an interface that exposes the minimal functionality required to
>> implement some of the popular compute APIs. This commit adds entry
>> points to set the grid layout and other state required to keep track
>> of the usual address spaces employed in compute APIs, to bind a
>> compute program, and execute it on the device.
>>
>> Reviewed-by: Marek Olšák <mar...@gmail.com>
>> ---
>> v2: Add "start slot" argument to the resource binding driver hooks.
>> v3: Split sampler views from shader resources.
>>
>> src/gallium/docs/source/context.rst | 39 +++++++++++++++
>> src/gallium/docs/source/screen.rst | 28 ++++++++++-
>> src/gallium/include/pipe/p_context.h | 73
>> ++++++++++++++++++++++++++++
>> src/gallium/include/pipe/p_defines.h | 19 +++++++-
>> src/gallium/include/pipe/p_screen.h | 12 +++++
>> src/gallium/include/pipe/p_shader_tokens.h | 9 ++++
>> src/gallium/include/pipe/p_state.h | 7 +++
>> 7 files changed, 185 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/gallium/docs/source/context.rst
>> b/src/gallium/docs/source/context.rst
>> index b2872cd..cb9b8de 100644
>> --- a/src/gallium/docs/source/context.rst
>> +++ b/src/gallium/docs/source/context.rst
>> @@ -542,3 +542,42 @@ These flags control the behavior of a transfer object.
>> ``PIPE_TRANSFER_FLUSH_EXPLICIT``
>> Written ranges will be notified later with :ref:`transfer_flush_region`.
>> Cannot be used with ``PIPE_TRANSFER_READ``.
>> +
>> +
>> +Compute kernel execution
>> +^^^^^^^^^^^^^^^^^^^^^^^^
>> +
>> +A compute program can be defined, bound or destroyed using
>> +``create_compute_state``, ``bind_compute_state`` or
>> +``destroy_compute_state`` respectively.
>> +
>> +Any of the subroutines contained within the compute program can be
>> +executed on the device using the ``launch_grid`` method. This method
>> +will execute as many instances of the program as elements in the
>> +specified N-dimensional grid, hopefully in parallel.
>> +
>> +The compute program has access to four special resources:
>> +
>> +* ``GLOBAL`` represents a memory space shared among all the threads
>> + running on the device. An arbitrary buffer created with the
>> + ``PIPE_BIND_GLOBAL`` flag can be mapped into it using the
>> + ``set_global_binding`` method.
>> +
>> +* ``LOCAL`` represents a memory space shared among all the threads
>> + running in the same working group. The initial contents of this
>> + resource are undefined.
>> +
>> +* ``PRIVATE`` represents a memory space local to a single thread.
>> + The initial contents of this resource are undefined.
>> +
>> +* ``INPUT`` represents a read-only memory space that can be
>> + initialized at ``launch_grid`` time.
>> +
>> +These resources use a byte-based addressing scheme, and they can be
>> +accessed from the compute program by means of the LOAD/STORE TGSI
>> +opcodes.
>> +
>> +In addition, normal texture sampling is allowed from the compute
>> +program: ``bind_compute_sampler_states`` may be used to set up texture
>> +samplers for the compute stage and ``set_compute_sampler_views`` may
>> +be used to bind a number of sampler views to it.
>> diff --git a/src/gallium/docs/source/screen.rst
>> b/src/gallium/docs/source/screen.rst
>> index 05f7e8f..5d8280a 100644
>> --- a/src/gallium/docs/source/screen.rst
>> +++ b/src/gallium/docs/source/screen.rst
>> @@ -110,7 +110,8 @@ The integer capabilities:
>> * ``PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY``: This CAP
>> describes
>> a hw limitation. If true, pipe_vertex_element::src_offset must always be
>> aligned to 4. If false, there are no restrictions on src_offset.
>> -
>> +* ``PIPE_CAP_COMPUTE``: Whether the implementation supports the
>> + compute entry points defined in pipe_context and pipe_screen.
>>
>>
>> .. _pipe_capf:
>> @@ -186,6 +187,29 @@ to be 0.
>> samplers.
>>
>>
>> +.. _pipe_compute_cap:
>> +
>> +PIPE_COMPUTE_CAP_*
>> +^^^^^^^^^^^^^^^^^^
>> +
>> +Compute-specific capabilities. They can be queried using
>> +pipe_screen::get_compute_param.
>> +
>> +* ``PIPE_COMPUTE_CAP_GRID_DIMENSION``: Number of supported dimensions
>> + for grid and block coordinates. Value type: ``uint64_t``.
>> +* ``PIPE_COMPUTE_CAP_MAX_GRID_SIZE``: Maximum grid size in block
>> + units. Value type: ``uint64_t []``.
>> +* ``PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE``: Maximum block size in thread
>> + units. Value type: ``uint64_t []``.
>> +* ``PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE``: Maximum size of the GLOBAL
>> + resource. Value type: ``uint64_t``.
>> +* ``PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE``: Maximum size of the LOCAL
>> + resource. Value type: ``uint64_t``.
>> +* ``PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE``: Maximum size of the PRIVATE
>> + resource. Value type: ``uint64_t``.
>> +* ``PIPE_COMPUTE_CAP_MAX_INPUT_SIZE``: Maximum size of the INPUT
>> + resource. Value type: ``uint64_t``.
>> +
>> .. _pipe_bind:
>>
>> PIPE_BIND_*
>> @@ -223,6 +247,8 @@ resources might be created and handled quite differently.
>> * ``PIPE_BIND_SCANOUT``: A front color buffer or scanout buffer.
>> * ``PIPE_BIND_SHARED``: A sharable buffer that can be given to another
>> process.
>> +* ``PIPE_BIND_GLOBAL``: A buffer that can be mapped into the global
>> + address space of a compute program.
>>
>> .. _pipe_usage:
>>
>> diff --git a/src/gallium/include/pipe/p_context.h
>> b/src/gallium/include/pipe/p_context.h
>> index 8b4a158..3c0b89e 100644
>> --- a/src/gallium/include/pipe/p_context.h
>> +++ b/src/gallium/include/pipe/p_context.h
>> @@ -63,6 +63,7 @@ struct pipe_vertex_element;
>> struct pipe_video_buffer;
>> struct pipe_video_decoder;
>> struct pipe_viewport_state;
>> +struct pipe_compute_state;
>> union pipe_color_union;
>> union pipe_query_result;
>>
>> @@ -141,6 +142,10 @@ struct pipe_context {
>> void (*bind_geometry_sampler_states)(struct pipe_context *,
>> unsigned num_samplers,
>> void **samplers);
>> + void (*bind_compute_sampler_states)(struct pipe_context *,
>> + unsigned start_slot,
>> + unsigned num_samplers,
>> + void **samplers);
>> void (*delete_sampler_state)(struct pipe_context *, void *);
>>
>> void * (*create_rasterizer_state)(struct pipe_context *,
>> @@ -220,6 +225,10 @@ struct pipe_context {
>> unsigned num_views,
>> struct pipe_sampler_view **);
>>
>> + void (*set_compute_sampler_views)(struct pipe_context *,
>> + unsigned start_slot, unsigned
>> num_views,
>> + struct pipe_sampler_view **);
>> +
>> void (*set_vertex_buffers)( struct pipe_context *,
>> unsigned num_buffers,
>> const struct pipe_vertex_buffer * );
>> @@ -418,6 +427,70 @@ struct pipe_context {
>> */
>> struct pipe_video_buffer *(*create_video_buffer)( struct pipe_context
>> *context,
>> const struct
>> pipe_video_buffer *templat );
>> +
>> + /**
>> + * Compute kernel execution
>> + */
>> + /*@{*/
>> + /**
>> + * Define the compute program and parameters to be used by
>> + * pipe_context::launch_grid.
>> + */
>> + void *(*create_compute_state)(struct pipe_context *context,
>> + const struct pipe_compute_state *);
>> + void (*bind_compute_state)(struct pipe_context *, void *);
>> + void (*delete_compute_state)(struct pipe_context *, void *);
>> +
>> + /**
>> + * Bind an array of buffers to be mapped into the address space of
>> + * the GLOBAL resource. Any buffers that were previously bound
>> + * between [first, first + count - 1] are unbound after this call.
>> + *
>> + * \param first first buffer to map.
>> + * \param count number of consecutive buffers to map.
>> + * \param resources array of pointers to the buffers to map, it
>> + * should contain at least \a count elements
>> + * unless it's NULL, in which case no new
>> + * resources will be bound.
>> + * \param handles array of pointers to the memory locations that
>> + * will be filled with the respective base
>> + * addresses each buffer will be mapped to. It
>> + * should contain at least \a count elements,
>> + * unless \a resources is NULL in which case \a
>> + * handles should be NULL as well.
>> + *
>> + * Note that the driver isn't required to make any guarantees about
>> + * the contents of the \a handles array being valid anytime except
>> + * during the subsequent calls to pipe_context::launch_grid. This
>> + * means that the only sensible location handles[i] may point to is
>> + * somewhere within the INPUT buffer itself. This is so to
>> + * accommodate implementations that lack virtual memory but
>> + * nevertheless migrate buffers on the fly, leading to resource
>> + * base addresses that change on each kernel invocation or are
>> + * unknown to the pipe driver.
>> + */
>> + void (*set_global_binding)(struct pipe_context *context,
>> + unsigned first, unsigned count,
>> + struct pipe_resource **resources,
>> + uint32_t **handles);
>> +
>> + /**
>> + * Launch the compute kernel starting from instruction \a pc of the
>> + * currently bound compute program.
>> + *
>> + * \a grid_layout and \a block_layout are arrays of size \a
>> + * PIPE_COMPUTE_CAP_GRID_DIMENSION that determine the layout of the
>> + * grid (in block units) and working block (in thread units) to be
>> + * used, respectively.
>> + *
>> + * \a input will be used to initialize the INPUT resource, and it
>> + * should point to a buffer of at least
>> + * pipe_compute_state::req_input_mem bytes.
>> + */
>> + void (*launch_grid)(struct pipe_context *context,
>> + const uint *block_layout, const uint *grid_layout,
>> + uint32_t pc, const void *input);
>> + /*@}*/
>> };
>>
>>
>> diff --git a/src/gallium/include/pipe/p_defines.h
>> b/src/gallium/include/pipe/p_defines.h
>> index 8b6d00d..c4c217b 100644
>> --- a/src/gallium/include/pipe/p_defines.h
>> +++ b/src/gallium/include/pipe/p_defines.h
>> @@ -304,6 +304,7 @@ enum pipe_transfer_usage {
>> #define PIPE_BIND_STREAM_OUTPUT (1 << 11) /*
>> set_stream_output_buffers */
>> #define PIPE_BIND_CURSOR (1 << 16) /* mouse cursor */
>> #define PIPE_BIND_CUSTOM (1 << 17) /* state-tracker/winsys
>> usages */
>> +#define PIPE_BIND_GLOBAL (1 << 18) /* set_global_binding */
>>
>> /* The first two flags above were previously part of the amorphous
>> * TEXTURE_USAGE, most of which are now descriptions of the ways a
>> @@ -346,7 +347,8 @@ enum pipe_transfer_usage {
>> #define PIPE_SHADER_VERTEX 0
>> #define PIPE_SHADER_FRAGMENT 1
>> #define PIPE_SHADER_GEOMETRY 2
>> -#define PIPE_SHADER_TYPES 3
>> +#define PIPE_SHADER_COMPUTE 3
>> +#define PIPE_SHADER_TYPES 4
>>
>>
>> /**
>> @@ -477,6 +479,7 @@ enum pipe_cap {
>> PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY = 65,
>> PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY = 66,
>> PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY = 67,
>> + PIPE_CAP_COMPUTE = 68
>> };
>>
>> /**
>> @@ -522,6 +525,20 @@ enum pipe_shader_cap
>> PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS = 18
>> };
>>
>> +/**
>> + * Compute-specific implementation capability. They can be queried
>> + * using pipe_screen::get_compute_param.
>> + */
>> +enum pipe_compute_cap
>> +{
>> + PIPE_COMPUTE_CAP_GRID_DIMENSION,
>> + PIPE_COMPUTE_CAP_MAX_GRID_SIZE,
>> + PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE,
>> + PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE,
>> + PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE,
>> + PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE,
>> + PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
>> +};
>>
>> /**
>> * Composite query types
>> diff --git a/src/gallium/include/pipe/p_screen.h
>> b/src/gallium/include/pipe/p_screen.h
>> index 45c441b..7ae7c9a 100644
>> --- a/src/gallium/include/pipe/p_screen.h
>> +++ b/src/gallium/include/pipe/p_screen.h
>> @@ -98,6 +98,18 @@ struct pipe_screen {
>> enum pipe_video_profile profile,
>> enum pipe_video_cap param );
>>
>> + /**
>> + * Query a compute-specific capability/parameter/limit.
>> + * \param param one of PIPE_COMPUTE_CAP_x
>> + * \param ret pointer to a preallocated buffer that will be
>> + * initialized to the parameter value, or NULL.
>> + * \return size in bytes of the parameter value that would be
>> + * returned.
>> + */
>> + int (*get_compute_param)(struct pipe_screen *,
>> + enum pipe_compute_cap param,
>> + void *ret);
>> +
>> struct pipe_context * (*context_create)( struct pipe_screen *,
>> void *priv );
>>
>> diff --git a/src/gallium/include/pipe/p_shader_tokens.h
>> b/src/gallium/include/pipe/p_shader_tokens.h
>> index df2dd5e..9d08fde 100644
>> --- a/src/gallium/include/pipe/p_shader_tokens.h
>> +++ b/src/gallium/include/pipe/p_shader_tokens.h
>> @@ -166,6 +166,15 @@ struct tgsi_declaration_resource {
>> unsigned ReturnTypeW : 6; /**< one of enum pipe_type */
>> };
>>
>> +/*
>> + * Special resources that don't need to be declared. They map to the
>> + * GLOBAL/LOCAL/PRIVATE/INPUT compute memory spaces.
>> + */
>> +#define TGSI_RESOURCE_GLOBAL 0x7fff
>> +#define TGSI_RESOURCE_LOCAL 0x7ffe
>> +#define TGSI_RESOURCE_PRIVATE 0x7ffd
>> +#define TGSI_RESOURCE_INPUT 0x7ffc
>> +
>> #define TGSI_IMM_FLOAT32 0
>> #define TGSI_IMM_UINT32 1
>> #define TGSI_IMM_INT32 2
>> diff --git a/src/gallium/include/pipe/p_state.h
>> b/src/gallium/include/pipe/p_state.h
>> index a459a56..74f4ebd 100644
>> --- a/src/gallium/include/pipe/p_state.h
>> +++ b/src/gallium/include/pipe/p_state.h
>> @@ -580,6 +580,13 @@ struct pipe_resolve_info
>> unsigned mask; /**< PIPE_MASK_RGBA, Z, S or ZS */
>> };
>>
>> +struct pipe_compute_state
>> +{
>> + const struct tgsi_token *tokens; /**< Compute program to be executed. */
>> + unsigned req_local_mem; /**< Required size of the LOCAL resource. */
>> + unsigned req_private_mem; /**< Required size of the PRIVATE resource. */
>> + unsigned req_input_mem; /**< Required size of the INPUT resource. */
>> +};
>>
>> #ifdef __cplusplus
>> }
>> --
>> 1.7.10
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev@lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
pgpQF3Wx0aS4T.pgp
Description: PGP signature
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
